CS 323 — Numerical analysis and computing, Spring 2025

Root finding: Secant method

Peng Zhang
2025-02-07
 Today’s plan
• Root-finding: secant method

• Reference: [Atkinson-Han] Chapter 3.3

 Root finding
• Given a function f : ℝ → ℝ, we want to find x* ∈ ℝ such that f(x*) = 0.

• x* is called a root or zero of f.

• Easy to solve analytically:

• f(x) = 3x - 2, f(x) = x2 + 2x - 3

• Hard to solve analytically:

• f(x) = x2 + ecos(x) - 3

• We have covered: the bisection method, Newton’s method

 Recap: Newton’s method
• Consider f : ℝ → ℝ with a root x*.
• We have an estimate of x*, denoted by x0,
which is assumed to be near x*.

• We approximate the graph y = f(x) with a

tangent line, and approximate the root x*
with the root of the straight line.

Root of f Our estimate of x*

x*
Root of the
tangent line
 Recap: Newton’s method
• Consider f : ℝ → ℝ with a root x*.
• We have an estimate of x*, denoted by x0,
which is assumed to be near x*.

• We approximate the graph y = f(x) with a

tangent line, and approximate the root x*
with the root of the straight line.

• Can we use a different straight-line Root of f Our estimate of x*

approximation?
x*
• This leads to the secant method. Root of the
tangent line
 Secant method
• Assume two initial guesses for x*: x0, x1.

• Two points (x0, f(x0)), (x1, f(x1)) determine a straight line, called the secant line.

• The secant line approximates the graph y = f(x), and its root x2 approximates x*.
Secant line Secant line

x2: root of the secant line

x*
x*
x2: root of the secant line
 Formula for x2
/
(Xo
f(x)
*

!
,

• The equation of the secant line: slope !

-
O
↓
O
define pla to b
e.

x*
x*
 Formula for x2 PIXzD =
f(X ) ,

+ (Xz -

Xi) .

If
• Formula for its root x2: Solving p(x2) = 0, we obtain

I
=
o

=>
(Xz-X . .
)
I
f(x
=

)
-

=> XX
X -X , = -

f(x) .
.

fat-fix

x*
x*
 Secant method
• Drop x0, use x1, x2 as a new set of approximate values for x*. This leads to an
improved value x3. Repeat this process.

• General iteration formula:

• This is called the secant method.

x*
 Secant method
• Drop x0, use x1, x2 as a new set of approximate values for x*. This leads to an
improved value x3. Repeat this process.

• General iteration formula:

• This is called the secant method.

• It is called a two-point method, since two

approximate values are needed to obtain an
improved value. Another example: the bisection
method. x*
 Secant vs Newton iteration
• Secant method:
xn − xn−1
E
xn+1 = xn − f(xn) ⋅
f(xn) − f(xn−1)
=
I(x)
f(xn 1)
[(xn)
Xn
-

-
Xn -
1
-

• Newton’s method:
t
f(xn)

C
xn+1 = xn − [Xn- f(xa 1))
f′(xn)
-

f(Xu))
f(xn) − f(xn)
For xn ≈ xn−1, f′(xn) ≈
-

• -

xn − xn−1
E
-
the line
slope of the slope of Jecant

line
tangent
Example 3.3.1 in [Atkinson-Han]
6
• Find a root of f(x) = x − x − 1 using the secant method.
Example 3.3.1 in [Atkinson-Han]
6
• Find a root of f(x) = x − x − 1 using the secant method.

• Choose x0, x1.

I xn − xn−1
For n ≥ X
0, apply the iteration: xn+1 = xn − f(xn) ⋅
• f(xn) − f(xn−1)
 *

* X

The secant method converges faster than the bisection method, but slower than Newton’s method.
The first few iterations don’t converge fast. But as xn becomes closer to x*, the convergence rate
increases.
 Convergence
• For the secant method to converge, we require the same assumption on f as
Newton’s method:

• f has two continuous derivatives for all x in some interval around x* (root)

• f′(x*) ≠ 0 (aka, the graph of y = f(x) is not tangent to the x-axis when the graph intersects it at x = x*; it
implies that x* is a simple root)

• Since f′ is continuous, f′(x) ≠ 0 for all x near x*

• The case f′(x*) = 0 is treated in [AH] Chapter 3.5 (won’t cover in class)
 Convergence
• Claim 1: Under the above assumption, for any n ≥ 1,

n−1 (
2f′(ζn) )
f′′(ξn)
• x* − xn+1 = (x* − xn )(x* − x ) ⋅ −

• where ζn is between xn−1, xn, and ξn is between the largest and the
smallest of xn−1, xn, x*.
 Convergence
• Claim 1: Under the above assumption, for any n ≥ 1,

n−1 (
2f′(ζn) )
f′′(ξn)
• x* − xn+1 = (x* − xn )(x* − x ) ⋅ −

• where ζn is between xn−1, xn, and ξn is between the largest and the
smallest of xn−1, xn, x*.

• Similar to the errors in Newton’s method:

( 2f′(cn) )
2 f′′(x*)
• x* − xn+1 = (x* − xn ) ⋅ −
 Convergence
• Claim 1: Under the above assumption, for any n ≥ 1, if xn, xn−1 are
sufficiently close to x*, then

n−1 (
2f′(x*) )
f′′(x*)
• x* − xn+1 ≈ (x* − xn )(x* − x ) ⋅ −
 Convergence
• Claim 1: Under the above assumption, for any n ≥ 1, if xn, xn−1 are
sufficiently close to x*, then Let M
=i
-
n−1 (
2f′(x*) )
f′′(x*)
• x* − xn+1 ≈ (x* − xn )(x* − x ) ⋅ −
- --

en + 1 en en-1
-

M
• Similar to the error in Newton’s method:

( 2f′(x*) )
2 f′′(x*)
• x* − xn+1 ≈ (x* − xn ) ⋅ −
2
-
enti - En +

M
 Convergence
f′′(x*)
Let M = − and en = x* − xn for n ≥ 0.
• 2f′(x*)

• en+1 ≈ enen−1M = Mento

(Men)(Ment)
• Similar to the error Newton’s method:

2
• en+1 ≈ en M => Menu (Mens
 Convergence
f′′(x*)
Let M = − and en = x* − xn for n ≥ 0.
• 2f′(x*)
Let Sn = Men . for n= o

• Men+1 ≈ (Men)(Men−1) => Sati SnOn-

• Similar to the error Newton’s method:

2
• Men+1 ≈ (Men) => Saxo S ??
 Convergence
f′′(x*)
Claim 2: Let δn = | − (x* − x ) | . Then, as xn → x*,
• 2f′(x*) n

r
1+ 5
δ
• n+1 ≈ δn, where r= ≈ 1.62 (the Golden ratio)
2
We know Sati SnSn-

Apply Snor = S = Suda =

So Sn-
UCr-1)
Apply Sn Sn = Sa -1
=
Sn-
=> Vr D vr
1 = 1
-

=
= 0
-
 Convergence
Example :

f′′(x*) Sn 10-5
Claim 2: Let δn = | − (x* − x ) | . Then, as xn → x*,
=

• 2f′(x*) n
secant :

r
1+ 5 Sny105X1 .
62

δ
• n+1 ≈ δn, where r= ≈ 1.62 (the Golden ratio)
-
2
Men + = Mehr
• Recall Newton’s method: Newton :
Recall
:

2 Sn *
Sni 10 5x2
• δn+1 ≈ δn
=
M(x -

Xn) =

• Newton’s method converges faster than the secant method.

 Order of convergence
• Let En := | x* − xn | (or En is a bound of | x* − xn | )

• Secant method: En+1 ≈M r−1

⋅ r
En r = 1 .
62

• Super linear convergence

2
• Newton’s method: En+1 ≈M⋅ En

• Quadratic convergence

1
• Bisection method: En+1 ≤ ⋅ En
2
• Linear convergence
 X
secant iteration : Xn += Xn-f(x) ·

Stopping criteria
• Stop the iterations when | xn+1 − xn | < tol (a pre-specified tolerance)
• By the mean-value theorem:
• f(xn) − f(x*) = f′(c)(xn − x*) for some c between x* and xn
f(xn) f(xn)
Since f(x*) = 0, we have x* − xn = − ≈ − = x − x
• f′(c) f′(x )
n+1 n
E n
IFX
-

***
 Assumption on Convergence
Methods Initial guess Each iteration Convergence
f rate

An interval (a,b)
Need evaluation Always
Bisection Continuous such that f(a) Linear
of f(x) convergence
f(b) < 0

Converge if f’, f’’

Continuous and Need evaluation exist and
Newton An x0 near x* Quadratic
differentiable of f(x) and f’(x) continuous, and
f’(x*) is nonzero

Converge if f’, f’’

Two points x0, Need evaluation exist and Super linear, but
Secant Continuous
x1 near x* of f(x) continuous, and not quadratic
f’(x*) is nonzero
 Claim
of 1 :

secant iteration :

Xn + Xn - f(Xn)
X
=
-

Def *
en =
X -

Xn for no
S

*
= Xn = X -

en

.-en)
=>
en =
en-f(x *_ en)

By Taylor remainder thi :

fee R

f(x* fix+ tf(e

=
+ e) x + e +

↓
between X *, **
X e
O

let e =
-en ,
-

en -

Approximate ford ~ fix ) *

=> enyen-1 fix * ) - en) + Efx (en)

*
·
 -
f(x
*
)) - e) -[fix *
)) -

en -) + =f x
ent]

e-en
=em enem -
Cent

-
enen :

__)